Comparison of the SLAM algorithms: Hangar experiments

advertisement
MATEC Web of Conferences 4 2 , 0 3 0 0 9 (2016 )
DOI: 10.1051/ m atecconf/ 2016 4 2 0 3 0 0 9
C Owned by the authors, published by EDP Sciences, 2016
Comparison of the SLAM algorithms: Hangar experiments
Mehmet Korkmaz
1
1,a
1
, Nihat Yılmaz , Akif Durdu
1
Selcuk University in Konya, TURKEY
Abstract. This study purposes to compare two known algorithms in an application scenario of simultaneous
localization and mapping (SLAM) and to present issues related with them as well. Mostly used SLAM algorithms
Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) are compared with respect to the point of
accuracy of the robot states, localization and mapping. Because of considering the most implementations in the
previous studies, the simulation environments are chosen as big as possible to provide reliable results. In this study,
two different hangar regions are tried to be simulated. According to the outcomes of the applications, UKF-based
SLAM algorithm has superior performance over the EKF-based one, apart from elapsed time.
1 Introduction
Autonomous mobile robots should perform the
behaviours or any given tasks with a high degree of
autonomy. It is vital for guiding the robot in a desired
manner. Therefore, a robot has to know answers of two
critical questions: “Where am I?”, “What does the world
look like?”. The first question addresses to the problem
which is known as a localization problem. According to
the localization problem, a mobile robot needs to know
its position (location, orientation) to achieve the goal [1].
On the other hand, the second question is related with the
environmental information. The robot seeks the answer of
what is the world around its. This problem is known as a
mapping phenomenon, in the literature. In this problem, a
robot tries to build a map or a plan of its environment
with an acquired sensor data [2].
In addition to the above mentioned parts, it is good to
specify that robot positions and environment information
are known for mapping and localization problem,
respectively. With these thought, simultaneous
localization and mapping (SLAM) problem addresses to
the problem that lacking of position and surrounding data.
Accordingly, SLAM by a mobile robot is concerned with
the problem of creating a map of an unknown
environment. The robot has its on-board sensors while
simultaneously determining its location using the created
map. Because of the fact that, truly autonomy system
needs to location and position of features of robot and
environment, SLAM problem solution is seen a “holy
grail” for the mobile robotic community [3]. The
background of SLAM problem is based upon the paper of
Smith et al. [4] by using stochastically approach to the
problem. After that, SLAM problem has attracted distinct
attention from the many researches and mobile robotics
communities in the past twenty five years [3,5].
a
From emerging to now SLAM has been found widely
usage area. Early studies, still ongoing, of SLAM are
performed by using EKF [6], then particle filter based
solutions are shown up [7], after that using of the
information-state approaches have come up [8]. Kalman
filter based solutions have been still remained on the
researchers’ agenda. On the other hand, visual SLAM
based approaches have been aroused attention from many
studies on this field [9].
For the solution of SLAM problem, plenty of
techniques has been applied, but among them Bayesian
based approaches have founded the most useful and
successful ones. According to the Bayesian method,
sensor data observed from the robot are expressed
thorough a conditional probability distribution. Then,
Bayesian filter uses these distributions as an input and
performs two phases that are prediction and correction
steps to solve the problem. In the prediction phase, robot
motion model is taken into consideration and next state of
the robot is predicted. After that, related observations are
taken by sensors and used to correct the states.
The structure of the paper is as follows: First, the
mathematical probabilistic definitions on SLAM problem
is described to show the bases of it in section II.
Thereafter, in section III, used robot dynamics, which is
based upon Ackermann Modell, are shown. Moreover,
EKF and UKF based SLAM algorithms’ foundation are
elucidated in depth and state equations have been
revealed. Finally, implementation of the algorithms is
realized using the thought simulation environments and
Root Mean Squared Error (RMSE) and true-estimated
paths are pointed out.
Corresponding author: mkorkmaz@selcuk.edu.tr
This is an Open Access article distributed under the terms of the Creative Commons Attribution License 4.0, which permits distribution, and reproduction in any medium, provided the original work is properly cited.
Article available at http://www.matec-conferences.org or http://dx.doi.org/10.1051/matecconf/20164203009
MATEC Web of Conferences
2 SLAM problem
SLAM is an exploration of environment while
simultaneously tracking the robot’s pose. In order to
overcome the problem, probabilistic approaches have
been applied successively, up to now. SLAM problem is
defined in two manners, which are online SLAM (eq. 1)
and full SLAM (eq. 2).
( , |: , : )
(: , |: , : )
(1)
(2)
Full SLAM can be distinguished from the online
SLAM only if the posterior are calculated over the entire
path : along with the map. In the equations, x
represents the location and orientation of the robot, is
a control vector, denotes the ith feature location and is an observation taken via robot’s sensor.
As a result of Bayes’ theorem and assumptions that
the robot motion model is Markov and the landmarks are
time invariant in the environment. With this thought,
SLAM formula can be reproduced as in the eq. 3 [10, 3].
1
,"
The equations above are summarized with the following
9
definitions: = 7 , , 8 robot pose information,
9
" = 7," , ," 8 location of the feature which is also
called landmark, = [ , ]9 control vector =
[$ , % ]9 observation. Where the equations symbols
means are identified as follows: velocity, robot
steering, $ observed range, % observed bearing, !
wheel base of the robot, time interval between ; and
; − 1.
The eq. 6 is called velocity motion model and the
other one, 7, represents the range-bearing observation
model (Fig. 1 and Fig. 2).
is a normalizing constant which comes from
assumptions stage. From the point of probabilistic view,
calculation of the probabilistic equations are in need of
motion and observation models that possesses the robot
(eq. 4-5). The motion model is obtained from the robot’s
dynamic equation while the observation one is derived
according to the sensor readings.
(4)
(5)
The eq. 3 tells that two stages are implemented to find
SLAM problem solution: time update (prediction), prior
state estimation and robot motion model are combined;
measurement update (correction), this calculation is
corrected and associated with the sensor observation and
finally result is normalized to appropriate interval.
Figure 1. Velocity motion model.
3 Robot model and Algorithms
In this section, Ackermann steering geometry that is used
for the robot model and EKF-UKF algorithms are
explained in detail.
3.1 Robot Model
In the study, used robot motion model has Ackermann
steering geometry. It is generally accepted having poor
kinematic model but also is a good representation for
slow motions [11, 12]. Following equations point out the
motion and observation models ((6), (7)), respectively.
+ ( + )
= = + ( + ) + ( )/!
(7)
5,3 54
( , | , ) =
. (
| , ) ∫ ( | , ) ( , | , ) (3)
( | , )
(
| , )
1
*-," − 0 + -," − 0
$,"
= #% & = '
6
2,3 24
,"
tan
− Figure 2. Range-bearing observation model.
(6)
03009-p.2
ICCMA 2015
3.2 EKF-based SLAM
Extended Kalman Filter (EKF) is the oldest and one of
the mostly used techniques to implement SLAM problem,
in the literature. EKF-based SLAM algorithm is a result
of Bayesian Filter and extension of Kalman Filter (KF)
for the form of nonlinear cases. The main goal of the
EKF is to find posterior distribution of the mean and
covariance ((8), (9)) by applying the prediction and
correction steps [13, 3, 14].
@|
̅ = ? = #
&
(8)
A
| = B
55
95C
5C
D
CC
(9)
|
In the time update or prediction step, robot motion
model and robot’s mean-covariance are used to predict
the mean and covariance, at time k-1. At this step, used
robot motion model ( ( | , )) is described in the
eq. 10.
= E( , ) + F
(10)
In the equation, E(∙) refers to robot’s kinematics and F
is an additive, zero mean Gaussian disturbance
(F ~H(0 , J )).
According to this step, predicted mean and covariance are
evaluated with the following equations ((11), (12)).
@| = E-@| , 0
55,| = ∇E55,| ∇E 9 + J
(11)
(12)
The most prominent part of the EKF is ∇E , which
means Jacobian of E calculated at time k-1.
The second stage is a measurement update or
correction that provides with the estimation of the
prediction via obtained measurements and measurement
model. Measurement model of the robot ((
| , ))
can be thought as an equation in 13.
5 = ℎ( , ) + M
(13)
In the equation h (∙) denotes the geometry of the
measurement, which is similar to motion model, and M is
an additive, zero mean Gaussian disturbance
(M ~H(0 , N ) ). Predicted measurement are calculated
from the (13) after that the important points that are
innovation and its covariance are obtained with respect to
real measurements and predicted values ((14), (15)).
= − ℎ(@| , A )
9
O = ∇h| ∇h + N
(14)
(15)
As in the ∇E, ∇h is a Jacobian of the ℎ evaluated at
preceding step. After from all these steps, estimation is
finished ((16), (17)), states are corrected and if necessary
next state can be repeated using the same steps.
B
@|
@|
D=B
D + P A |
A |
| = | − P O P9
(16)
K is a Kalman gain and calculated as the following (18).
P = | ∇h9 O
(18)
3.3 UKF-based SLAM
Unscented Kalman Filter (UKF) is an alternative way to
implement SLAM problem. UKF is mainly based on
unscented transform. The main idea on the UKF is to
express the probability distribution with random variables,
so called sigma points. Instead of using nonlinear
function to define probability distribution by Taylor
expansion, it is chosen 2 + 1 sigma points from the
Gaussian (19).
R
R
["]
["]
R [S] = ̅ = ?
= ̅ + -T ( + U)0" , E$ = 1, … , (19)
= ̅ − -T ( + U)0" , E$ = + 1, … ,2
where λ is defined,
U = 1( + ;) − (20)
α (usually set to 1V − 4 ≤ ≤ 1 ) and k (generally
chosen 0) are the scaling factors that show the spreads of
sigma points from the mean.
In the light of unscented transform information,
mentioned sigma points are replacement in the motion
model (21) and predicted mean and covariance are
calculated according to the equations (22), (23),
respectively.
"
"
R|
= E-R|
, 0
(21)
"
1Z
@| = ∑"\S F" R|
(22)
9
"
"
1Z
| = ∑"\S F" 7R| − @| 87R| − @| 8 + J
(23)
In the correction step, sigma points are substituted in the
measurement (observation) model ((24), (25)).
"
"
|
= ℎ-R|
0
"
1Z
̂| = ∑"\S F" |
= − ̂|
(24)
(25)
(26)
9
"
"
O = ∑1Z
"\S F" 7| − ̂| 87| − ̂| 8 + N (27)
(26) is called innovation and (27) is an innovation
covariance. After the measurement, corrected mean and
covariance at time k-1 are evaluated with respect to (16)
and (17) as in the same EKF-based SLAM. In the above
equations, K is a Kalman gain (28) and is 5_ cross
correlation matrix (29).
P = 5_,| O
5_,|
(28)
9
"
"
= ∑1Z
@| 87|
− ̂| 8 (29)
"\S F" 7R| − As it is seen in the formulas UKF-based SLAM
differs from the EKF one that is not include Jacobian
calculation which generally provides an advantage of
easy calculation [15, 16].
(17)
03009-p.3
MATEC Web of Conferences
4 Simulation results
The proposed SLAM algorithms both EKF and UKF are
simulated using the MATLAB program. In the simulation,
the robot dynamic is thought as Ackermann model. The
robot wheel base is set to 4 meters, speed is 3 m/s. In
addition to this, sensor maximum distance is accepted as
a 30 meters and time interval, ∇t, between k and k-1 is
supposed to be 0.025 seconds.
In the study, two different scenarios are implemented
that are hangar_1 (square low) and hangar_2 (cycloid).
The simulated maps are indicated in the figures, 3 and 4.
According to the simulation, hangar environments are
nearly chosen in the size of 200 by 200 meters to see the
effectiveness of the algorithms in a big size region. It is
approximately placed a special feature, called also
landmark, in each ten meters.
For the fair comparison, each algorithm is run ten
times and average RMSE criteria results are indicated in
the table 1 and 2, for each experiment.
Table 1. Hangar_1 (square low) RMSE and Time Results.
Algorithm
Position
error
Heading
error
Time
(s)
EKF
4,43432
0,78978
62,1163
UKF
2,55331
0,60402
72,1636
Table 2. Hangar_2 (cycloid) RMSE and Time Results.
Algorithm
Position
error
Heading
error
Time
(s)
EKF
3,63777
1,78697
0,54790
0,20794
46,0049
49,6039
UKF
Figure 5 and 6 are the results of the algorithms’
estimated path versus true robot path. In addition to
previous results these figures also show the supremacy of
the UKF-based algorithm, which is the closest one to the
true path.
Figure 3. Hangar_1 (square low) simulation environment
Figure 5. Hangar_1 (square low) true path vs algorithm’s
estimated path
Figure 4. Hangar_2 (cycloid) simulation environment
It is used a root mean square error (RMSE) criteria to see
the productivity of the algorithms for both position (30)
and heading angle, subtracting from true data to
estimated one.
N`Ob = *
c
∑e
@ 3 )c
3f(53 5@3 ) d(23 2
g
(30)
where, " and " are a true paths and @" and @" are an
estimated paths by algorithms, N is an number of state
obtained in simulation.
Figure 6. Hangar_2 (cycloid) true path vs algorithm’s estimated
path
03009-p.4
ICCMA 2015
5 Conclusion
EKF is one of the best and classical algorithm to the
solution of SLAM problem. Although its easy
implementation and effectiveness are verified various
studies, new solution to SLAM problem are required.
Besides this, UKF is one of the mostly used techniques
and powerful solution to the SLAM problem. In this
study, both EKF and UKF algorithms are applied to the
big size environment to see the effectuality of the
algorithms. First of all, simulation environments are
generated taking consideration to the square low and
cyclical hangar plans. After that, landmarks are
established nearly each ten meters and robot waypoints
are determined to optimal vision of sensor. Finally,
produced data files are loaded to the workspace and
algorithms are tested navigating the robot through
waypoints. Both position error and heading error
evaluation results show the superiority of the UKF to
EKF (table 1, 2). While the RMSE of the robot pose is
superior from the point of the UKF, elapsed time to
complete the route is not as good as in EKF.
Acknowledgements
References
2.
3.
4.
5.
6.
7.
S. Thrun, Y. Liu, D. Koller, A. Ng, and H. DurrantWhyte, “Simultaneous localisation and mapping with
sparse extended information filters,” Int. J. Robot.
Res., 23 (7–8), pp. 693–716, (2004).
9.
J. Fuentes-Pacheco, J. Ruiz-Ascencio, & J.M.
Rendón-Mancha, Visual simultaneous localization
and mapping: a survey. Artificial Intelligence
Review, 43(1), 55-81, (2015)
C.C. Wang, C. Thorpe, & S. Thrun, Online
simultaneous localization and mapping with
detection and tracking of moving objects: Theory
and results from a ground vehicle in crowded urban
areas. In Robotics and Automation, 2003.
Proceedings.
ICRA'03.
IEEE
International
Conference on, 1, pp. 842-849). IEEE, (2003,
September)
S. Lavalle, Planning Algorithms. Cambridge:
Cambridge University Press, (2006)
L. Marin, M. Vallés, A. Soriano, A. Valera, & P.
Albertos, Event-Based Localization in Ackermann
Steering Limited Resource Mobile Robots,
IEEE/ASME Trans. Mechatronics, 19(4), 1171-1182
(2014)
A. Chatterjee, A. Rakshit, & N.N. Singh,
Simultaneous Localization and Mapping (SLAM) in
Mobile Robots. In Vision Based Autonomous Robot
Navigation (pp.
167-206).
Springer
Berlin
Heidelberg, (2013)
P.S. Maybeck, Stochastic Models, Estimaton and
Control, vol. I. New York: Academic 1979.
S. Li, & P. Ni, Square-root unscented Kalman filter
based simultaneous localization and mapping.
In Information and Automation (ICIA), 2010 IEEE
International Conference on (pp. 2384-2388). IEEE,
(2010, June)
R. Havangi, M.A. Nekoui, H.D. Taghirad, & M.
Teshnehlab, SLAM based on intelligent unscented
Kalman filter. In The 2nd International Conference
on Control, Instrumentation and Automation, (2011)
10.
11.
12.
13.
This work is supported by Selcuk University
Scientific Research Fund, BAP. Authors are thankful to
RAC-LAB (http://www.rac-lab.com/) for providing the
trial version of their commercial software.
1.
8.
14.
15.
Z, Liang, X Ma, & X. Dai, Extended Monte Carlo
algorithm to collaborate distributed sensors for
mobile robot localization. In Robotics and
Biomimetics, 2007. ROBIO 2007. IEEE International
Conference on (pp. 1647-1652), IEEE, (2007,
December)
S. Thrun, Robotic mapping: A survey. Exploring
artificial intelligence in the new millennium, 1-35,
(2002)
H. Durrant-Whyte, & T. Bailey, Simultaneous
localization and mapping: part I, IEEE Robot.
Autom. Mag. 13(2), 99-110, (2006)
R. Smith, M. Self, and P. Cheeseman. A stochastic
map for uncertain spatial relationships. In
International Symposium of Robotics Research,
pages 467–474, (1987)
T. Bailey, J. Nieto, J. Guivant, M. Stevens, & E.
Nebot, Consistency of the EKF-SLAM algorithm.
In Intelligent Robots and Systems, 2006 IEEE/RSJ
International Conference on (pp. 3562-3568). IEEE,
(2006, October)
R. Smith and P. Cheesman, “On the representation of
spatial uncertainty,” Int. J. Robot. Res., 5(4), pp. 56–
68, (1987)
G. Dissanayake, H.F. Durrant-Whyte and Tim Bailey,
“A Computationally Efficient Solution to the
Simultaneous Localization and Map Building
Problem”, In IEEE Int. Conf. on Robotics and
Automation, 2, pp. 1009-1014, (2000).
16.
03009-p.5
Download